Add `wp.bfloat16` scalar data type

[shi-eric]

Summary

Add bfloat16 (brain floating-point 16) as a first-class scalar type in Warp, on par with the existing wp.float16. The bfloat16 format uses 8 exponent bits and 7 mantissa bits, giving it the same dynamic range as float32 at half the memory, which makes it widely used in ML training and inference workloads.

Scope

Python type system

• New wp.bfloat16 scalar type registered alongside existing float types
• Round-trip conversions between float32 and bfloat16 (round-to-nearest-even, with correct NaN handling)
• Direct array construction from Python floats, lists, tuples, and NumPy float32 data
• Support in vectors, matrices, and structs

Native C++/CUDA

• wp_bfloat16 struct with float/half conversions and arithmetic operators
• Platform-specific conversion paths for CUDA (__CUDA_ARCH__), Clang, and generic C++
• Comparison operators with float promotion for correct IEEE 754 NaN/zero semantics

Kernel support

• Arithmetic, comparison, and casting builtins
• Math function overloads (abs, sqrt, log, exp, trig functions, etc.)
• Autodiff support
• Atomic operations: add, min, max (with CAS fallback)
• Tile operations including matmul via cuBLASDx

Framework interop

• DLPack import/export
• PyTorch torch.bfloat16 tensors (zero-copy where possible)
• JAX jax.numpy.bfloat16 arrays
• Fabric and sparse array support

Compile-time optimization

• WP_NO_BFLOAT16 guard in builtin.h, cuda_crt.h, and tile.h so modules that don't use bfloat16 skip compiling its overloads, reducing LLVM codegen time